From Deep Learning to Digital Insights: Twenty-Five Years of Understanding How Students Learn in Higher Education

Introduction

This journal celebrates its 25th anniversary as its very first issue was published in July 2000. The launch of the journal also marked the start of my own academic career. Hence, I accepted the editor-in-chiefs’ invitation to write a contribution for the anniversary special issue in which I share my “reflections on gains made in active learning research and current gaps and challenges for engaging and supporting students in their learning” also as a way to look back to my own research in the past 25 years and share my personal view on the challenges ahead for the coming decades.

From the start of my academic journey, I—like many other researchers and practitioners—have been interested in how we can enhance student learning in higher education. As was common among scholars 25 years ago, my early research focused on understanding the relationship between key characteristics of the learning environment and how students learn in higher education. Already in the very first issue of this journal, Rosie (2000) explored a similar question by examining whether students’ use of a dialectical approach could promote “deep learning.” The notion that higher education should foster deep approaches to learning has appeared repeatedly in publications since the journal’s inception. This view is largely rooted in the idea that higher education was—and continues to be—expected to prepare work-ready graduates capable of lifelong learning and, given its scholarly nature, should cultivate deep learning strategies (Asikainen & Gijbels, 2017; Lake & Boyd, 2015). As a result, many researchers, myself included, have sought to examine how features of the learning environment relate to students’ deep (or surface) approaches to learning.

In the remainder of this article, I begin by reflecting on the origins of the concepts of deep and surface approaches as used in active learning research in higher education, and on how these concepts have been measured. I then outline how research in this field has evolved over the past 25 years—from mainly cross-sectional, correlational studies investigating links between teaching–learning environment characteristics and students’ learning approaches, to more longitudinal research aimed at understanding whether students develop deeper approaches to learning during and across courses in higher education. Although this shift toward longitudinal research represents methodological progress, I argue that the field has long relied almost exclusively on self-report measures. I then describe how, over the last decade, the increasing focus on learning supported by digital tools has enabled researchers to move beyond self-reports by combining them with other, often more behavioral, measures such as clickstream data or eye-tracking. I conclude by presenting an example of an ongoing research project that demonstrates how integrating the capabilities of artificial intelligence with the measurement tools used over the past 25 years can yield new instruments that empower higher education teachers to better support their students’ active learning processes.

Deep and Surface Approaches to Learning

The concepts of deep and surface approaches to learning were first introduced by Ference Marton in 1976 to describe how students engage with their studies in everyday academic contexts, drawing on both experimental results and interview data. In their early work, Marton and Säljö (1976a, 1976b) examined processing strategies specifically in the context of reading academic texts. These studies found that students’ processing strategies were shaped by their perceptions of the teaching–learning environment, rather than reflecting their general study habits—a misconception occasionally found in later interpretations (Richardson, 2015). Building on this foundational research, other scholars later developed instruments such as Biggs’s (1987) Study Process Questionnaire (SPQ) and Entwistle and Ramsden’s (1983) Approaches to Studying Inventory (ASI) to investigate approaches to learning, as well as related constructs like Vermunt’s learning patterns (e.g. Vermunt & Donche, 2017), in broader contexts.

These questionnaires typically distinguish between (at least) two qualitatively different approaches to learning: deep and surface (Biggs, 2003; Marton, 1976; Marton & Säljö, 1984; Prosser & Trigwell, 1999). A deep learning approach reflects students’ aims to truly comprehend the material and engage meaningfully with it. They concentrate on core concepts and underlying principles, applying strategies that help them construct real understanding. The surface approach to learning, on the other hand, refers to students selectively memorizing, based on motives or intentions that are extrinsic to the real purpose of the task, such as fear of failure or keeping out of trouble (Vanthournout et al., 2014).

Relating Learning Environments to How Students Learn

Since the earliest studies by Marton and Säljö, it has been clear that the teaching–learning environment plays a significant role in shaping students’ approaches to learning. More specifically, students’ perceptions and experiences of this environment are strongly linked to the approaches they adopt (Ramsden, 1997). Even 25 years after the launch of Active Learning in Higher Education, understanding this relationship remains a “hot topic,” as demonstrated by the journal’s lists of most-read and most-cited articles in recent years (e.g. Hailikari, et al., 2021; Helker et al., 2024). Nonetheless, the research landscape on this theme has evolved considerably.

At the beginning of this century, most studies were cross-sectional and relied on data gathered through student interviews (e.g. Rosie, 2000) or through self-report questionnaires that assessed both how students learn (e.g. the SPQ) and how they perceive the learning and assessment environment (e.g. Gijbels & Dochy, 2006). Without going into every detail of these findings here, the general conclusion was that positive perceptions of the teaching–learning environment tend to be associated with a deep approach to learning, while negative perceptions are more often linked to a surface approach (Asikainen & Gijbels, 2017; Entwistle et al., 2003; Parpala, et al., 2010; Richardson, 2005, 2006).

In the discussion sections of many of these cross-sectional and largely correlational studies—including my own—researchers consistently pointed out the limitations of this type of research, particularly regarding questions of causality (i.e. what influences what?). One way to address these limitations was to shift toward more longitudinal research.

How do Students Evolve During Higher Education?

The belief that higher education fosters students’ development toward deeper approaches to learning is deeply embedded in our understanding of what higher education should accomplish. This assumption has motivated researchers to shift from cross-sectional studies to more longitudinal research conducted within and across higher education courses. Together with Henna Asikainen from the University of Helsinki, I conducted a systematic review of the available longitudinal studies across disciplines up to 2016 to examine how students’ approaches to learning evolve throughout their studies (Asikainen & Gijbels, 2017). However, based on the review’s findings, it was not possible to confirm that students indeed develop deeper approaches to learning during higher education. Instead, we concluded that there is no clear empirical evidence supporting the assumption that deep learning naturally increases over the course of higher education (Asikainen, 2014). Importantly, we argued that the lack of overall change in students’ approaches does not necessarily imply that no change occurs at all; rather, changes may take place within specific subgroups of students or under particular learning conditions. Haggis (2003) had previously suggested that fostering the deep approach may be extremely difficult if it is not already present in the student. Conversely, McCune and Entwistle (2011) proposed that some students possess a natural inclination to seek understanding, resulting in consistently high levels of deep learning across courses due to their strong motivation to fully grasp the subject matter. Richardson (2011) further argued that the stability of students’ approaches could reflect the similarity of their learning environments throughout higher education. Overall, the common claim that deep approaches are more stable than surface approaches was not supported by our review. In fact, more studies reported no significant change in surface approaches than in deep approaches. Notably, studies employing person-oriented instead of variable-oriented longitudinal methods (e.g. Fryer, 2017; Vanthournout et al., 2013) indicate that students who begin their studies with a deep approach tend to maintain it. The person-oriented studies included in our review consistently showed that different subgroups of students follow different developmental trajectories, and that these trajectories are highly individual (Fryer, 2017; Lindblom-Ylänne et al., 2013; Postareff et al., 2014, 2015; Saravanamuthu & Yap, 2014). This may also help explain why group-level studies produce inconsistent findings. Based on our review (Asikainen & Gijbels, 2017), we suggest that research on the development of learning approaches in higher education should move beyond the group level and focus more closely on individual patterns and subgroup differences.

A similar conclusion emerged from another systematic literature review that I conducted in collaboration with colleagues in the Netherlands. In this review (Dolmans et al., 2016), we focused specifically on how students’ approaches to learning change within one particular learning environment and disciplinary context in higher education. We reviewed studies in the health sciences that investigated students’ approaches to learning in problem-based learning settings. Our findings suggested that problem-based learning fosters deeper approaches to learning, while having relatively little effect on reducing surface approaches.

In both reviews, we noted a critical limitation: all included studies relied exclusively on self-reported data, did not triangulate these measures with alternative forms of evidence. Over the past decade, however, this situation has advanced. Increasingly, researchers have invested in integrating self-report measures with more behavioral forms of data, leading to the rise of multimodal research designs within the field of active learning in higher education. In the following section, I will describe how our research group in Antwerp has taken up this challenge and why I believe this development—especially in light of emerging artificial intelligence (AI) technologies—holds substantial promise for practitioners aiming to support student learning in higher education.

Multimodal Research

Over the past decade, we have successfully conducted multimodal studies in which we combined rich self-reported data—gathered through interviews or questionnaires—with newer types of data that allow us to capture students’ learning in a non-intrusive manner while they are actively engaged in learning tasks. These studies aimed to broaden our understanding of how students approach their learning when working on specific instructional activities. Given the prominence of particular learning formats in higher education, one series of studies focused on learning from instructional videos (e.g. Gijsen et al., 2024), while another concentrated on learning from text (e.g. Catrysse et al., 2016).

In designing these studies, we drew inspiration from the seminal research conducted by Marton et al. (1975). In their original experiments, students were asked to read three texts and prepare to answer questions about the content. The questions administered after the first two texts served as the only indication of how the students should approach reading the third text. Students in the deep-processing condition received questions requiring higher-level understanding (e.g. constructing a summary statement), whereas students in the surface-processing condition received reproduction-oriented questions. After studying the third text, the researchers conducted semi-structured interviews to examine how the manipulation had influenced the students’ processing strategies. Their findings suggested that students adapted their level of processing as intended (Marton, 1975; Richardson, 2000), demonstrating that processing depth can be influenced through targeted prompts or questions. This work also showed that the level of processing depends strongly on the expected form of assessment (Richardson, 2000). As noted earlier in this article, these foundational studies contributed to the later development of the concepts of deep and surface approaches to learning.

In our own very first studies on learning from text (Catrysse et al., 2016), we sought to replicate this original experiment by asking students to read three texts on a screen while we used a screen-based eye tracker to capture their reading behavior. After each text, students received questions in a manner similar to the original design. The key difference from the earlier research was the use of eye-tracking technology to record students’ eye movements during reading. After the third text, we conducted a stimulated-recall interview in which students were asked to explain their learning behavior, using a replay of their gaze patterns as the stimulus. Based on the analyses of the eye-movement data and the students’ verbalizations, we found that students in the deep condition did not necessarily spend more time looking at the essential parts of the text than students in the surface condition; rather, they processed these elements in a deeper way. Additionally, students in the surface condition spent more time looking at factual details and reported repeating these facts more frequently. Follow-up studies provided further and more nuanced evidence that eye-tracking data can offer valuable insights into how students learn while completing different types of tasks, what strategies they use, and which strategies tend to be more successful (e.g. Catrysse et al., 2018, 2022).

While these findings are highly promising from a research perspective, one might question their practical value for teaching in higher education, given that analyzing eye-tracking data is both time-consuming and resource-intensive, limiting its direct applicability in classroom settings. Although this concern was certainly valid when we began conducting these studies a decade ago, the rise of AI has since created exciting new opportunities for integrating eye-tracking technologies into educational practice—both in today’s classrooms and even more so in the near future.

Future Directions

Eye-tracking research has revealed subtle patterns in learners’ attention and comprehension, offering deeper insight into how students interact with (digital) content (e.g. Catrysse et al., 2025; Mézière et al., 2023). These findings mark a shift: with the integration of AI, we can now move beyond observation toward real-time support. Without interrupting students with interviews or questionnaires, AI combined with observational data—such as eye-tracking—can provide immediate insight into how students learn and where they encounter difficulties. At the same time, new AI-driven tools create opportunities to assist teachers in making informed instructional decisions based on these insights.

This vision underpins the ongoing EYE-TEACH project, a collaboration among researchers and practitioners from nine European countries (www.eyeteach.eu). The project began from the observation that although researchers can derive valuable insights from eye-tracking data—for example, how students learn from expository texts—translating these findings into practical guidance for teachers remains challenging. Reading comprehension is a good case in point. Teachers traditionally monitor reading comprehension by observing behavior or eliciting students’ thinking processes, but these methods are time-consuming and difficult to apply consistently. Consequently, many teachers in higher education rely on product-oriented assessments, which reveal little about students’ actual reading processes (Leslie & Caldwell, 2017). Furthermore, teachers—also in higher education—often struggle to accurately assess reading levels and adapt instruction to individual needs (Knoop-van Campen et al., 2021).

The EYE-TEACH project aims to address this gap by equipping teachers with AI-enhanced eye-tracking tools that generate actionable insights into student learning—insights that were previously inaccessible. Such AI-supported approaches open new possibilities for capturing the dynamics of reading through collaboration between teachers and technology. Advances in AI in education now support not only routine tasks but also learner monitoring and the identification of appropriate interventions (Chen et al., 2020; Wang et al., 2024). This aligns with the augmentation principle—AI should enhance, not replace, teachers’ expertise (Cukurova et al., 2019) —and resonates with Human-Centered AI and hybrid intelligence perspectives that emphasize combining human and machine strengths (Dellermann et al., 2019; Molenaar, 2022; Shneiderman, 2022).

Eye-tracking can illustrate the potential of this synergy: by capturing fine-grained eye-movement data, it makes students’ reading processes visible and supports more adaptive forms of instruction (Carter & Luke, 2020; de la Peña, 2024; Mézière et al., 2023). However, the sheer volume and complexity of such data pose major challenges for real-time interpretation by teachers, particularly in the large classes common in many higher education programs. Within the EYE-TEACH project, AI is designed to act as a collaborative partner by analyzing eye-tracking data and translating it into meaningful, actionable guidance—such as real-time instructional adjustments, targeted interventions, or automated adaptations—presented to teachers in various formats. The tool is currently being developed in co-creation with educators. During the first workshops with higher-education teachers (see, e.g. Buseyne, 2025), teachers expressed a clear need for a system that provides both individual- and class-level data and agreed on the value of comparing how different students engage with the same material, supported by indicators such as engagement, attention, reading pace, patterns, and predicted understanding. They also showed interest in tools that could identify the most challenging parts of a text and recommend suitable interventions. The preferred way of receiving information typically involved visualizations paired with textual summaries, as well as spreadsheets and graphs for deeper analysis. Most teachers favored dashboards that combine the possibility to monitor students, being alerted by the tool when problems arise or getting support from the tool to make data-driven decisions. Importantly, teachers emphasized that the tool must preserve their autonomy; indeed decisions about students should always remain in the hands of teachers, new tools can however support teachers in making their decisions more informed and based on insights in students learning processes that teachers did not have before (Buseyne, 2025).

By briefly highlighting the aims of EYE-TEACH, I close this article with a hopeful view of how research and educational innovation can reinforce each other—precisely the mission of the journal celebrating its 25th anniversary in this special issue. It is clear that new challenges and ethical concerns lay in front of us and require again new methodologies and intensified collaborations between researchers and teachers in higher education. I am confident that the future of Active Learning in Higher Education is bright, and I hope to play a small role to contributing to it over the next 25 years.